DistillSleep: Real-Time, On-Device, Interpretable Sleep Staging from Single-Channel EEG.

IF 4.9 2区医学 Q1 Medicine

Sleep Pub Date : 2025-08-22 DOI:10.1093/sleep/zsaf240

Keondo Park, Joopyo Hong, Wooseok Lee, Hyun-Woo Shin, Hyung-Sin Kim

{"title":"DistillSleep: Real-Time, On-Device, Interpretable Sleep Staging from Single-Channel EEG.","authors":"Keondo Park, Joopyo Hong, Wooseok Lee, Hyun-Woo Shin, Hyung-Sin Kim","doi":"10.1093/sleep/zsaf240","DOIUrl":null,"url":null,"abstract":"Study objectives: Polysomnography (PSG) is the current gold standard for sleep staging but requires laboratory equipment, multiple sensors, and labor-intensive manual scoring. We developed DistillSleep, a single-channel electroencephalogram (EEG) framework that delivers accurate, real-time, and interpretable sleep staging on resource-constrained devices.Methods: DistillSleep consists of (1) a high-capacity teacher model and (2) a 109 k-parameter student model designed for edge deployment. Both incorporate a Multi-Wavelength Pyramid module and Transformer-based architecture to capture intra- and inter-epoch features. Feature- and prediction-level knowledge distillation transfers the teacher's expertise to the student. Training and evaluation used >10 000 overnight recordings from six cohorts (SHHS1, PhysioNet 2018, DCSM, KISS, SleepEDF-78, ISRUC), following AASM guidelines. Performance was assessed with Macro-F1.Results: The teacher achieved state-of-the-art Macro-F1 scores (SHHS1 81.1%, PhysioNet 78.9%, DCSM 81.2%, KISS 80.0%) and provided frequency-resolved saliency maps, inter-epoch context and well-calibrated confidence (ECE 0.07). The student maintained competitive accuracy (up to 79.7% Macro-F1) while executing <10 ms per 30-second epoch on three embedded platforms (Raspberry Pi 4B, Jetson orin nano, Coral dev board), reducing computational load 115-fold versus the best prior method (SleePyCo). Interpretability was transferred intact to the student, offering clinicians frequency-band importance and inter-epoch context visualizations, and calibration was further improved by 2.7$\\times$.Conclusions: DistillSleep combines expert-level accuracy, millisecond-scale latency, and transparent decision logic in a single-channel EEG form factor. These capabilities pave the way for point-of-care diagnostics, same-night therapy titration, and large-scale home monitoring, expanding the reach of sleep medicine while retaining clinical trust.","PeriodicalId":22018,"journal":{"name":"Sleep","volume":" ","pages":""},"PeriodicalIF":4.9000,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Sleep","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1093/sleep/zsaf240","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"Medicine","Score":null,"Total":0}

引用次数: 0

Abstract

Study objectives: Polysomnography (PSG) is the current gold standard for sleep staging but requires laboratory equipment, multiple sensors, and labor-intensive manual scoring. We developed DistillSleep, a single-channel electroencephalogram (EEG) framework that delivers accurate, real-time, and interpretable sleep staging on resource-constrained devices.

Methods: DistillSleep consists of (1) a high-capacity teacher model and (2) a 109 k-parameter student model designed for edge deployment. Both incorporate a Multi-Wavelength Pyramid module and Transformer-based architecture to capture intra- and inter-epoch features. Feature- and prediction-level knowledge distillation transfers the teacher's expertise to the student. Training and evaluation used >10 000 overnight recordings from six cohorts (SHHS1, PhysioNet 2018, DCSM, KISS, SleepEDF-78, ISRUC), following AASM guidelines. Performance was assessed with Macro-F1.

Results: The teacher achieved state-of-the-art Macro-F1 scores (SHHS1 81.1%, PhysioNet 78.9%, DCSM 81.2%, KISS 80.0%) and provided frequency-resolved saliency maps, inter-epoch context and well-calibrated confidence (ECE 0.07). The student maintained competitive accuracy (up to 79.7% Macro-F1) while executing <10 ms per 30-second epoch on three embedded platforms (Raspberry Pi 4B, Jetson orin nano, Coral dev board), reducing computational load 115-fold versus the best prior method (SleePyCo). Interpretability was transferred intact to the student, offering clinicians frequency-band importance and inter-epoch context visualizations, and calibration was further improved by 2.7$\times$.

Conclusions: DistillSleep combines expert-level accuracy, millisecond-scale latency, and transparent decision logic in a single-channel EEG form factor. These capabilities pave the way for point-of-care diagnostics, same-night therapy titration, and large-scale home monitoring, expanding the reach of sleep medicine while retaining clinical trust.

查看原文本刊更多论文

蒸馏睡眠：实时、设备上、可解释的单通道脑电图睡眠分期。

研究目的：多导睡眠图（PSG）是目前睡眠分期的黄金标准，但需要实验室设备、多个传感器和劳动密集型的人工评分。我们开发了DistillSleep，这是一个单通道脑电图（EEG）框架，可以在资源受限的设备上提供准确、实时和可解释的睡眠分期。方法：DistillSleep由(1)一个高容量教师模型和(2)一个为边缘部署设计的109 k参数学生模型组成。两者都采用了多波长金字塔模块和基于变压器的架构来捕获历元内和历元间的特征。特征和预测级别的知识升华将教师的专业知识传递给学生。训练和评估使用来自6个队列（SHHS1、PhysioNet 2018、DCSM、KISS、sleeppedf -78、ISRUC）的bb10万份夜间录音，遵循AASM指南。使用Macro-F1评估性能。结果：该教师达到了最先进的Macro-F1分数（SHHS1 81.1%, PhysioNet 78.9%, DCSM 81.2%, KISS 80.0%），并提供了频率分辨显著性图、历元间背景和校准良好的置信度（ECE 0.07）。在执行结论时，该学生保持了具有竞争力的准确性（高达79.7% Macro-F1）： DistillSleep在单通道EEG形式因素中结合了专家级别的准确性，毫秒级的延迟和透明的决策逻辑。这些功能为即时诊断、当晚治疗滴定和大规模家庭监测铺平了道路，在保持临床信任的同时扩大了睡眠医学的范围。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Sleep Medicine-Neurology (clinical)

CiteScore

8.70

自引率

10.70%

发文量

期刊介绍： SLEEP® publishes findings from studies conducted at any level of analysis, including: Genes Molecules Cells Physiology Neural systems and circuits Behavior and cognition Self-report SLEEP® publishes articles that use a wide variety of scientific approaches and address a broad range of topics. These may include, but are not limited to: Basic and neuroscience studies of sleep and circadian mechanisms In vitro and animal models of sleep, circadian rhythms, and human disorders Pre-clinical human investigations, including the measurement and manipulation of sleep and circadian rhythms Studies in clinical or population samples. These may address factors influencing sleep and circadian rhythms (e.g., development and aging, and social and environmental influences) and relationships between sleep, circadian rhythms, health, and disease Clinical trials, epidemiology studies, implementation, and dissemination research.